DMTree: Towards Efficient Tree Indexing on Disaggregated Memory via Compute-side...
summary
Prototype DMTree implements a compute-side collaborative B‑tree: leaf nodes keep key/value entries and per‑leaf fingerprint tables whose addresses live in internal nodes, while fingerprints and lock metadata are distributed across compute servers (round‑robin or consistent hashing) and cached with primary/replica sync and version validation. Evaluated on a 7‑node setup (6 compute, 1 memory) with YCSB/West‑store workloads, it reduces memory‑side RDMA LPS/bandwidth bottlenecks and yields up to 5.7× throughput vs. prior DM range indexes.
AI-written summary · about
chapters (13)
- 0:00DMTree intro and disaggregated memory
- 1:22Tree indexing challenges in IDM
- 2:49Trade-offs in IDM index designs
- 4:34Resource imbalance in IDM
- 5:38DMTree compute-side design
- 6:24Key locating with fingerprint tables
- 7:27Fingerprint storage offloading
- 8:09Compute-side fingerprint cache
- 9:33Distributed locking for bottlenecks
- 10:15DMTree optimizations
- 10:36Evaluation setup and workloads
- 10:57Performance results
- 12:02Conclusion and resources
related talks
14mSMART: A High-Performance Adaptive Radix Tree for Disaggregated Memory
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
15mOne-sided RDMA-Conscious Extendible Hashing for Disaggregated Memory
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
12mDirect Access, High-Performance Memory Disaggregation with DirectCXL
15mCitron: Distributed Range Lock Management with One-sided RDMA
19mFineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained...